Speed Control and Scheduling of Data Mules in Sensor Networks

Transcription

1 Speed Control and Scheduling of Data Mules in Sensor Networks RYO SUGIHARA and RAJESH K. GUPTA University of California, San Diego Unlike traditional multihop forwarding among stationary sensor nodes, use of mobile devices for data collection in wireless sensor networks has recently been gathering more attention. The use of mobility significantly reduces the energy consumption at sensor nodes, elongating the functional lifetime of the network. However, a drawback is an increased data delivery latency. Reducing the latency through optimizing the motion of data mules is critical for this approach to thrive. In this paper, we focus on the problem of motion planning, specifically, determination of speed of the data mule and the scheduling of the communication tasks with the sensors. We consider three models of mobility capability of data mule to accommodate different types of vehicles. Under each mobility model, we design optimal and heuristic algorithms for different problems: single data mule case, single data mule with periodic data generation case, and multiple data mules case. We compare the performance of the heuristic algorithm with a naive algorithm and also with multihop forwarding approach by numerical experiments. We also compare one of the optimal algorithms with a previously proposed method to see how our algorithm improves the performance and is also useful in practice. As far as we know, this study is the first of a kind that provides a systematic understanding of the motion planning problem of data mules. Categories and Subject Descriptors: C.2.1 [Computer-Communication Networks]: Network Architecture and Design Wireless communication General Terms: Algorithm, Design, Performance Additional Key Words and Phrases: Controlled mobility, motion planning, scheduling, linear programming, simulation 1. INTRODUCTION Controlled mobility presents an attractive alternative to multihop forwarding for efficient data collection in a sensor field. In particular, we consider collecting data from stationary sensor nodes using data mules via wireless communication. A data mule is a mobile node with radio and sufficient amount of storage to store the data from the sensors in the field. Data mules have been used in recent sensor network applications, e.g., a robot in underwater environmental monitoring [Vasilescu et al. 2005] and a UAV (unmanned aerial vehicle) in structural health monitoring [Todd et al. 2007]. A data mule travels across the sensor field and collects data from each sensor node while the distance is short, and later deposits all the data to the base station. In this way, each sensor node can conserve energy, since it only needs... Permission to make digital/hard copy of all or part of this material without fee for personal or classroom use provided that the copies are not made or distributed for profit or commercial advantage, the ACM copyright/server notice, the title of the publication, and its date appear, and notice is given that copying is by permission of the ACM, Inc. To copy otherwise, to republish, to post on servers, or to redistribute to lists requires prior specific permission and/or a fee. c 20YY ACM /20YY/ $5.00 ACM Journal Name, Vol. V, No. N, Month 20YY, Pages 1 28.

2 2 R. Sugihara and R. K. Gupta to send the data over a short distance and has no need to forward other sensors data all the way to the base station. Note that energy issue is critical for sensor nodes as opposed to the data mule that returns to the base station after the travel. However, one disadvantage of this approach is that it generally takes more time to collect data, which in turn incurs larger data delivery latency. Thus optimizing the data delivery latency is vital for the data mule approach to be useful in practice. Optimizing the motion of data mule is a hard problem in general, since we need to find a trajectory satisfying both spatial constraints (i.e., wireless communication range) and temporal constraints (i.e., time required for data collection), where the spatial constraints alone is analogous to traveling salesman problem. To deal with the hardness, previous studies on data mule approaches [Kansal et al. 2004; Somasundara et al. 2004; Ma and Yang 2006; Xing et al. 2007] simplified the problem by using simple models for mobility and communications. These simplifications lead to suboptimal solutions with unnecessarily large latency. Moreover, since these algorithms are often tailored for specific settings, it is difficult to compare them and use them for similar problems in different application scenarios. To address these limitations, in this paper we describe a problem framework that we call data mule scheduling (DMS) problem. The DMS problem captures the motion planning problem as one composed of loosely connected subproblems of path selection, speed control, and job scheduling. This framework enables us to identify the cases of the problem that we can optimally solve, as well as the hard cases that we explore heuristics. Our focus in this paper is on the one dimensional case of the DMS problem (1-D DMS). The 1-D DMS problem is to determine the speed change and data collection schedule to minimize the data delivery latency. This applies to many cases in which the data mules move on the fixed paths. We first discuss the basic case, where a single data mule collects a given amount of data from each sensor node. We show we can give an optimal speed change and data collection schedule when the data mule s mobility model is either constant speed or variable speed. For a case when there is an acceleration constraint, we design a heuristic algorithm. Later we discuss the cases of periodic data generation and multiple data mules. In the periodic data generation case, each sensor node generates data at a certain rate and the data mule travels repetitively. In the multiple data mules case, data mules move either along an identical path or arbitrary paths. For both of these cases, we design optimal and heuristic algorithms in a similar way as the basic case. By numerical experiments, the performance of the proposed algorithms is evaluated against multihop forwarding approach and a previously proposed speed control algorithm. Our contributions are: Formulate the data mule scheduling (DMS) problem, a problem framework for optimal control of data mule for minimizing the data delivery latency; Present optimal and heuristic algorithms for one dimensional case of the DMS problem (1-D DMS) in various problem settings including single data mule, multiple data mules, and periodic data generation, each for three different mobility models. The rest of this paper is organized as follows. In Section 2 we give an overview

3 Speed Control and Scheduling of Data Mules in Sensor Networks 3 node B Communication range node C Location job A B C Execution time e(a) e(b) e(c) Location Job A B C Execution time e(a) e(b) e(c) Time node A Speed Time A B C Time (a) Path selection (b) Speed control (c) Job scheduling Fig. 1. Subproblems of data mule scheduling of the DMS problem framework. We also introduce an application example and related work. In Section 3 we define the 1-D DMS problem in more details and introduce mobility models we consider. We also discuss the feasibility testing by processor demand analysis that we use in designing the optimal algorithms. We present optimal and heuristic algorithms for the basic cases of single data mule in Section 4. Periodic case with single data mule is discussed in Section 5. In Section 6, we discuss multiple data mules cases. Section 7 shows the results of the simulation experiments and Section 8 concludes the paper. 2. PROBLEM FRAMEWORK FOR MOTION PLANNING OF DATA MULE In this section we present the data mule scheduling (DMS) framework for optimizing the motion of data mule. As we have discussed in the introduction, motion planning of data mule is a hard problem. Communications with sensor nodes need to take place in the proximity of each node and will take certain time duration, whereas the motion of data mule is possibly governed by dynamics constraints. There is also a prioritization problem when the data mule is in the communication ranges of multiple nodes. To deal with this complexity, in the DMS framework, we decompose the problem into the following three subproblems as shown in Figure 1: (1) Path selection: which trajectory the data mule follows (2) Speed control: how the data mule changes the speed while moving along the path (3) Job scheduling: from which sensor the data mule collects data at each time point Path selection is to determine the trajectory of the data mule in the sensor field. To collect data from each particular sensor, the data mule needs to enter the sensor s communication range at least once. Depending on the capability of data mule, there may be some constraints on path selection, such as minimum turning radius. Speed control is the second subproblem to determine how the data mule changes its speed along the chosen path. The data mule needs to change the speed so that

4 4 R. Sugihara and R. K. Gupta it stays within each sensor s communication range long enough to collect all the data from it. Final subproblem is job scheduling. Once the time-speed profile is determined, we get a mapping from each location to a time point. Thus we get a scheduling problem by regarding data collection from each sensor as a job. Each job has one or more intervals in which it can be executed. Job scheduling is to determine the allocation of time to jobs so that all jobs can be completed. The DMS problem is general and can be used to express several earlier problems in the area. For instance, the assumption of no wireless communication (as in [Somasundara et al. 2004; 2007]) is easily expressed by setting the communication range to zero in the path selection subproblem. The constant speed assumption (as in [Ma and Yang 2006; 2007; Xing et al. 2007]) and variable speed assumption (as in [Zhao and Ammar 2003; Kansal et al. 2004]) are handled in the speed control subproblem. In this paper, we focus on the subproblems of speed control and job scheduling. These subproblems constitute the one dimensional case of data mule scheduling problem (1-D DMS). The 1-D DMS problem is important in many cases when the data mule needs to move along a given path. A solution for the 1-D DMS problem is twofold. One is time-speed profile, which determines the speed changes of the data mule. With a time-speed profile, each point on the location axis can be mapped onto a point on the time axis. Then we obtain a scheduling problem having a set of jobs, each of which has an execution time and feasible intervals. For this problem, we need to determine job schedule that defines when the data mule communicates with each node. The objective of the 1-D DMS problem is to find a speed control plan and a feasible job schedule so that the total travel time of the data mule is minimized. 2.1 Why Do We Minimize the Latency? As mentioned earlier, data mules can be used as an alternative to multihop forwarding in sensor networks. The use of data mules in collecting data introduces the trade-off between energy consumption and data delivery latency. Our goal is to optimize this trade-off through either minimizing the energy consumption under some latency constraint or minimizing the latency under some constraint on energy consumption. Protocol designers have tried to optimize the multihop forwarding in both energy and latency through sophisticated MAC protocols [Ye et al. 2002; Polastre et al. 2004; Rhee et al. 2005]. Data mule, or its combination with multihop forwarding, is a relatively nascent area. In this paper, we focus on the pure data mule approach, in which each node uses only direct communication with the data mule and no multihop forwarding. Energy consumption related to communication is already minimized in this case, since each node only sends its own data and does not forward others data. Naturally, our objective is to minimize the data delivery latency by minimizing the travel time of the data mule. 2.2 Example Application: SHM with UAV Our problem formulation is based on our experience with the example application described in [Todd et al. 2007]. It is a structural health monitoring (SHM)

5 Speed Control and Scheduling of Data Mules in Sensor Networks 5 application to do post-event (e.g., earthquakes) assessments for large-scale civil infrastructure such as bridges. Automated damage assessment using sensor systems is much more efficient and reliable than human visual inspections. In this application, the sensor nodes operate completely passively and do not carry batteries, for the sake of long-term measurement and higher maintainability. Upon data collection, an external mobile element provides energy to each node via microwave transmission, wakes it up, and collects data from it. The prototype system uses a radio-controlled helicopter as the mobile element that is either remotelypiloted or GPS-programmed. Each sensor node is equipped with ATmega128L microcontroller, a 2.4GHz XBee radio, antennas for transmission/reception, and a supercapacitor to store the energy. Each node has two types of sensors. One is a piezoelectric sensing element integrated with nuts and washers to check if the bolt has loosened. The other is capacitive-based sensors for measuring peak displacement and bolt preload. Since the size of data from these sensors are small, communication time is almost negligible; however, it takes a few minutes to charge a supercapacitor through microwave transmission in the current prototype. The team is currently investigating a new design to improve the charging time down to tens of seconds. The data collected by the UAV is brought back to the base station and analyzed by researchers using statistical techniques for damage existence and its location/type. Since the primary purpose of this application is to assess the safety of large civil structures after a disaster such as an earthquake, every process including data collection and analysis needs to be as quick as possible for prompt recovery. Furthermore, shorter travel time is required in view of the limited fuel on the helicopter. Thus the goal of our formulation is to achieve data collection from spatially distributed wireless sensors in the minimum amount of time. It also provides another reason for using data mule approach instead of multihop forwarding: simply because the SHM sensors are not capable of doing multihop communication. Furthermore, use of UAVs implies the need for more precise mobility model that takes acceleration constraint into consideration, as opposed to the simple move or stop model used in majority of the related work. 2.3 Related Work Data mule approaches, or using controlled mobility for improving data collection performance in sensor or mobile ad hoc networks, have been discussed in various papers from both theoretical and applications perspectives [Zhao and Ammar 2003; Ho and Fall 2004; Vasilescu et al. 2005; Ma and Yang 2006; Somasundara et al. 2006; Somasundara et al. 2007; Xing et al. 2008; Chebrolu et al. 2008]. For an overview of data mule and similar approaches, see [Ekici et al. 2006]. Our work on the DMS problem has been focused on identifying the combinatoric nature of the problem among its various variants in an attempt to develop a comprehensive understanding of fundamental performance limits of data mule approach. Ma and Yang [2006; 2007] study the problem of maximizing the network lifetime, which is defined as the time until the first node dies (i.e. minimum of the lifetime of all nodes), by finding an optimal motion of a data mule. Speed control is not their focus and they just use a constant speed mobility model. They also avoid

6 6 R. Sugihara and R. K. Gupta the problem of scheduling by assuming either that data communication time is negligible (in [Ma and Yang 2006]) or that the data mule stops while communicating with sensor nodes (in [Ma and Yang 2007]). However, the former assumption does not apply when data size is big and communication is slow, and the latter results in inefficiency when the data mule can actually communicate while it is moving. Zhao and Ammar [2003] study the use of controlled mobility in mobile ad hoc networks. They assume a controllable mobile node (called ferry ) that mediates communications between sparsely deployed stationary nodes. The speed of ferry is basically constant but can be reduced when it is necessary to communicate more data with a node. They have formulated the problem as a linear programming problem. Our work is similar and we build our formulations upon theirs, but with extensions that include more general mobility models, as well as other scenarios including multiple data mules and periodic motions. A more sophisticated speed controlling scheme has been proposed by Kansal et al. [2004] and Somasundara et al. [2006]. Assuming a data mule periodically travels across the sensor field along a fixed path, they propose an adaptive speed control algorithm. In the algorithm, the data mule slows down when it encounters the nodes from which data collection has not been very successful in the previous period. Their speed control algorithm is reactive in the sense it is solely based on the past performance. By contrast, we study a proactive approach when we have a priori information on the communication range and data collection time. For multiple data mules case, Jea et al. [2005] study the case in which multiple data mules move on fixed paths. They present a distributed coordination scheme for allocating sensor nodes to each data mule to achieve a good distribution of communications load. They just assume movements at the given constant speed. Our work is complementary to their work and focused on identifying the performance limits of centralized schemes when we can freely determine the path and speed of each data mule. This would serve as a performance target for distributed schemes. Our formulation as the DMS problem is closely related to scheduling. It is known that the Earliest Due Date (EDD) rule, in which the earliest deadline among all available jobs is executed at any time slice, is optimal for single processor preemptive scheduling with release times [Stankovic et al. 1995]. On the other hand, our problem is not standard in that each job may have multiple feasible intervals. As far as we know, there are only few studies on this case: for unit-length, nonpreemptible jobs [Simons and Sipser 1984] and for preemptible, non-continuable (i.e., jobs must be completed within one feasible interval) jobs [Shih et al. 2003; Chen et al. 2005], both of which are different from our problem of interest D DMS PROBLEM In this section, we define the 1-D DMS problem in a more formal way. We first introduce some terms and definitions and then give a formal definition. We also discuss the mobility models of data mules. 3.1 Terminology, Definitions, and Assumptions For the job scheduling subproblem, a job τ i has an execution time e i and a set I i of feasible intervals. A feasible interval I I i is a time interval [r(i),d(i)], where r(i) is a release time and d(i) is a deadline. A job can be executed only within its

7 Speed Control and Scheduling of Data Mules in Sensor Networks 7 feasible intervals. A simple job is a job with one feasible interval, whereas a general job can have multiple feasible intervals. For instance, in Figure 1(c), job B and C are simple jobs and job A is a general job. Similarly for the speed control subproblem, a location job τ i has an execution time e i and a set I i of feasible location intervals. A feasible location interval I I i is a location interval [r(i),d(i)], where r(i) is a release location and d(i) is a deadline location. A location job can be executed only within its feasible location intervals. A simple location job is a location job with one feasible location interval, whereas a general location job can have multiple feasible location intervals. In Figure 1(b), location job B and C are simple location jobs and location job A is a general location job. For an interval I = [r,d] (also for a location interval), I denotes the length d r. We also define containment as follows: I I if and only if r r and d d where I = [r,d ]. We make the following assumptions. Each sensor node is stationary and the location is known. Communication range and execution time are known. Communication is always successful in the communication range. All location jobs are preemptible without any cost incurred and can be executed over multiple feasible location intervals. There is no dependency among the location jobs. Data mule can communicate with one node at a time. Depending on the dynamics constraint, data mule may have constraints on the maximum speed and maximum acceleration. 3.2 Mobility Models In the speed control subproblem, we consider three different constraints on the dynamics of the data mule. The first is constant speed, where the data mule cannot change the speed after it starts to move, i.e., v(t) = v 0 for some constant v 0. The second is variable speed, where the data mule can instantaneously change the speed within a speed range. i.e., v min v(t) v max for given constants v min,v max. The third is variable speed with acceleration constraint model. In this model, the data mule can change the speed, but the rate of change is within the maximum absolute acceleration, i.e., dv(t)/dt a max for given constant a max. The constant speed and variable speed models apply to ground vehicles such as Packbot 1, which is commonly used as a data mule in actual deployments. The acceleration constrained model captures mobility more precisely and is most appropriate when we cannot ignore the inertia, for example in case that a helicopter is used as a data mule as in [Todd et al. 2007]. 3.3 Problem Definition Instance of the 1-D DMS problem is (L, J ), where [0,L]: total travel interval of the data mule on the location axis J : set of location jobs; i-th location job τ i is characterized by I i : set of feasible location intervals 1

8 8 R. Sugihara and R. K. Gupta e i : execution time Solution to the problem is a pair of time-speed profile v(t) and job schedule. Let T denote the total travel time. Then the constraints on the motion of data mule are T v(t) = L and the dynamics constraints according to each mobility model. 0 After v(t) is determined, we can define a function f(x) that maps location x to time t. Using f(x), we obtain a job scheduling problem, which we call an induced job scheduling problem. We need to determine v(t) such that the induced job scheduling problem has a feasible schedule : a job schedule that finishes all the jobs before their deadlines. The objective of the 1-D DMS problem is to find a solution that minimizes T. 4. ALGORITHMS FOR BASIC CASES In this section, we consider the 1-D DMS problem for single data mule case. We first introduce a feasibility test by processor demand analysis. Using that, we design optimal algorithms for constant speed case and variable speed case. Then we present the idea of heuristic algorithm for the acceleration constrained case. 4.1 Preliminary: Feasibility Testing by Processor Demand Analysis To design optimal algorithms for the 1-D DMS problem, we use processor-demandbased feasibility testing by Baruah et al. [1993]. Processor demand g in interval [t 1,t 2 ] is the sum of the execution time of the tasks whose feasible interval is completely contained in the interval, and is defined as g(t 1,t 2 ) = e(τ). τ J, I(τ) [t 1,t 2 ] Then the following theorem holds for periodic tasks with arbitrary relative deadline (i.e., relative deadline of each task can be smaller than its period): Theorem 4.1 [Baruah et al. 1993]. Let τ = {T 1,...,T n } be a task system. τ is not feasible iff there exist natural numbers t 1 < t 2 such that g(t 1,t 2 ) > t 2 t 1 Using the theorem, we can show that testing at each release time and deadline is sufficient for guaranteeing the feasibility. In other words, Theorem 4.2. Task system is feasible iff g(t 1,t 2) t 2 t 1 for any t 1 {r i },t 2 {d i } satisfying t 1 < t 2. The proof is in the appendix. This feasibility testing is for periodic task system in real time scheduling, but we can apply it for our case. Specifically, we can determine the speed for each location interval so that the processor demand for the interval is at most the time that the data mule stays in the interval. 4.2 Constant Speed For constant speed case, the problem is to find the maximum speed v 0 such that all jobs can be finished. We present two optimal offline scheduling algorithms, each for simple location jobs and general location jobs, respectively.

9 Speed Control and Scheduling of Data Mules in Sensor Networks Simple Location Jobs. When each location job has one feasible location interval, following simple algorithm in Algorithm 1 finds the maximum possible v 0 such that all location jobs can be completed. It applies processor-demand based feasibility test (Theorem 4.2) for all possible pairs of a release location and a deadline location. Algorithm 1 Find-Min-MaxSpeed 1: for each location interval I = [r(τ ),d(τ )] s.t. τ,τ J, r(τ ) d(τ ) do 2: d = e(τ) Processor demand for I τ J, I(τ) I 3: u[i] I Maximum speed allowed for I d 4: end for 5: return min I u[i] This algorithm runs in O(n 3 ) time where n is the number of location jobs in a naive implementation, but we can improve it to O(n 2 ) by computing the processor demand incrementally. Specifically, for each starting location, by having a list of jobs sorted by their deadline locations, we can incrementally extend the interval and calculate the processor demand in O(1) time. Then it takes O(n) time for each starting location, and since there are at most n starting locations, it takes O(n 2 ) as a whole General Location Jobs. For general location jobs case, we formulate the problem as a linear programming problem. We split the total travel interval [0,L] into (2m+1) location intervals [l 0 (= 0),l 1 ], [l 1,l 2 ],...,[l 2m,l 2m+1 (= L)] (l i l i+1 ), where m is the number of feasible location intervals of all location jobs, and each l i is either a release location or a deadline location. For each location job τ J, we consider variables p 0 (τ),...,p 2m (τ), in which p i (τ) represents the time allocated to job τ within the location interval [l i,l i+1 ]. Let v 0 denote the speed of data mule. Then we have the following linear program: Maximize v 0 Subject to 0 i 2m, τ J, p i (τ) 0. (Feasible intervals) 0 i 2m, τ J, if [l i,l i+1 ] I(τ), p i (τ) = 0. (Job completion) τ J, 2m i=0 p i(τ) = e(τ). (Processor demand) 0 i 2m, τ J p i(τ) (l i+1 l i )/v 0. The processor demand constraint becomes a linear constraint by introducing a new variable u 0 = 1/v 0. For the obtained solution, we can make a job schedule in the following way. Location interval [l i,l i+1 ] is mapped to the time interval [ i 1 k=0 z k, i k=0 z k], where z k = (l k+1 l k )/v 0. For each time interval, we allocate p i (τ 1 ) for location job τ 1 from the start of the interval, p i (τ 2 ) for τ 2 after that, and continue this for all location jobs.

10 10 R. Sugihara and R. K. Gupta 4.3 Variable Speed In variable speed case, the data mule can change its speed anytime. To make the problem realistic 2, we enforce constraints on speed for this case and the data mule can choose its speed within the range [v min,v max ] for given v min and v max Simple Location Jobs. For simple location jobs case, we have the following optimal offline algorithm shown in Algorithm 2. The algorithm is based on processor demand analysis and uses Find-Min-MaxSpeed internally. Interestingly, this algorithm is equivalent to the one for optimal dynamic voltage scaling presented in [Yao et al. 1995]. Algorithm 2 Sequential-Find-Min-MaxSpeed 1: loop 2: v c Find-Min-MaxSpeed 3: if v c < v min then return INFEASIBLE 4: else if v c > v max then 5: Set v max for all remaining intervals and finish 6: else 7: Set v c for the current tight interval 8: Remove jobs within the tight interval 9: Compress remaining jobs 10: end if 11: end loop Here is how this algorithm works. For the set of location jobs, Find-Min- MaxSpeed finds a tight interval and the corresponding speed v c. It is the minimum speed that makes the processor demand for each location interval equal or less than the time allocated to it. In other words, if the data mule moves at the speed faster than v c, there is at least one interval in which the allotted time is less than the processor demand (thus violates the feasibility). Therefore, if v c < v min (Line 3), it is infeasible. Compress (Line 9) is an operation used in [Yao et al. 1995], which is to remove the tight interval from the feasible intervals of all jobs and to connect the remaining two intervals together to construct a new set of jobs. In each iteration, v c is nondecreasing. This is shown by the same reasoning in [Yao et al. 1995]. When v c reaches v max, we cannot increase the speed of any remaining location intervals. Thus we set the speed to v max for all these intervals (Line 5). The optimality of this algorithm is shown immediately by using the proof in [Yao et al. 1995]. This algorithm runs in O(n 3 ) time, since each iteration takes O(n 2 ) time by using the improved implementation of Find-Min-MaxSpeed and at least one job is removed at each iteration. As for online algorithms, when v min = 0, the following algorithm shown in Algorithm 3 is an optimal online algorithm that minimizes the total travel time. In the algorithm, EDD-with-Stop, the data mule moves at v max while executing a job 2 Without speed constraints, the data mule can always minimize the total travel time simply by moving at infinite speed and stopping to execute a job (and repeat this for each job).

11 Speed Control and Scheduling of Data Mules in Sensor Networks 11 having the earliest deadline, just in the same way as ordinary Earliest Due Date algorithm. However, when a job is not completed at its deadline, the data mule stops until the job is completed and moves at v max again. Algorithm 3 EDD-with-Stop Init J P : set of pending location jobs, i.e., the location jobs that are currently executable and not finished yet, init with v: data mule s speed, init with v max a(τ): time allocated to job τ, init with 0 x c : current location, init with X s On τ J P, x c = d(τ) When a job is unfinished at its deadline location 1: J u {τ τ J P,d(τ) = x c } Set of jobs that needs to be finished here 2: v 0 Data mule stops 3: Complete each job in J u 4: J P J P \ J u 5: v v max Move at v max again 6: τ ed arg min τ JP d(τ) Job with the earliest deadline location 7: Execute τ ed On τ J, x c = r(τ) 8: J P J P {τ r(τ) = x c } 9: τ ed arg min τ JP d(τ) 10: Execute τ ed On τ J P, a(τ) = e(τ) 11: J P J P \ τ 12: τ ed arg min τ JP d(τ) 13: if τ ed then Execute τ ed 14: end if When jobs released When jobs finished We have a following theorem about the optimality of EDD-with-Stop. The proof is in the appendix. Theorem 4.3. EDD-with-Stop is optimal for the 1-D DMS problem for simple location jobs with variable speed model when v min = General Location Jobs. For general location jobs, we design an offline algorithm by linear programming formulation. The formulation is similar to the one in constant speed case (Section 4.2.2), but we have additional variable z i for each location interval [l i,l i+1 ] to represent the time that the data mule spends in the interval. The linear program is as follows: Minimize 2m i=0 z i Subject to 0 i 2m, τ J, p i (τ) 0, z i 0. (Feasible intervals) 0 i 2m, τ J, if [l i,l i+1 ] I(τ), p i (τ) = 0. (Job completion) τ J, 2m i=0 p i(τ) = e(τ). (Processor demand) 0 i 2m, τ J p i(τ) z i.

12 12 R. Sugihara and R. K. Gupta Speed Speed Tight interval Speed Location Accel Plateau interval Decel Recursively maximize interval interval (a) (b) (c) Fig. 2. Idea of the heuristic algorithm: (a) Increase the plateau speed. Dotted curves show the acceleration / deceleration for the fastest possible travel covering the whole interval. Bold and thin lines mean fixed intervals and free intervals, respectively. (b) A tight interval is found. (c) Recursively maximize the speed for the remaining free intervals. (Max/min speed) 0 i 2m, v min z i l i+1 l i v max z i. 4.4 Variable Speed with Acceleration Constraint When there is a constraint on acceleration, the problem is NP-hard for general location jobs case 3 [Sugihara and Gupta 2007]. Here we briefly present a heuristic algorithm that gives a feasible solution for both simple and general location jobs cases. For more details on the algorithm, please refer to [Sugihara and Gupta 2010] Approach. Figure 2 shows the idea of the heuristic algorithm. The algorithm works recursively, and in each recursion, we confine ourselves to the following 3-phase speed changing profile: first accelerate at the maximum acceleration, then move at the constant speed, and finally decelerate at the maximum negative acceleration. We call each of these intervals accel interval, plateau interval, and decel interval, respectively. Further we call the speed in the plateau interval as plateau speed. As shown in Figure 2(b), the main idea of the algorithm is to maximize the plateau speed until we have a tight interval, which is defined as an interval whose length (in time) is equal to the processor demand for that interval. We can naturally extend the definition to define tight location interval, when we give the speed of data mule for the interval. For the intervals in the plateau interval but not in the tight interval, we can still increase the speed without violating the feasibility. As shown in Figure 2(c), we recursively apply the maximization procedure to them until there is no such interval Details. The heuristic algorithm consists of following four steps: Step 1. Simplify: Converting general location jobs to simple location jobs. For each general location job, distribute the execution time to each feasible location interval proportionally to its length. Step 2. Maximize: Find maximum plateau speed and tight interval. In a similar way as Sequential-Find-Min-MaxSpeed algorithm, for each interval of release location of deadline location, calculate the plateau speed that makes it a tight interval. Maximum plateau speed is the minimum of these plateau speeds. From Theorem 4.2, this procedure does not destroy the feasibility, since the processor 3 Complexity is unknown for simple location jobs case.

13 Speed Control and Scheduling of Data Mules in Sensor Networks 13 demand is equal or less than the time given for each location interval of a release location and a deadline location. Step 3. Trim: Trim feasible location intervals of each location job. We first process the accel interval and the decel interval, and then the tight interval. For each job, we calculate the time that needs to be allocated in the remaining location interval after trimming. We use the EDD algorithm and the LRT (latest release time) algorithm [Liu 2000], which is equivalent to the EDD in reversed time axis and is optimal. Step 4. Recursion: By the previous step, all remaining jobs are separated into two disjoint groups. One group of jobs populate the location interval from the end of the accel interval to the beginning of the tight interval, and the other group populate the location interval from the end of the tight interval to the beginning of the decel interval. The speed for these intervals are not fixed yet, i.e., there may still be some room for increasing the speed without destroying the feasibility. Thus, we recursively maximize the speed by repeating from Step 2 for these intervals. The number of recursions varies from zero to two depending on the configuration of the tight interval. 5. PERIODIC 1-D DMS PROBLEM In the periodic case of the 1-D DMS problem, each sensor node generates data at a given rate and a data mule travels across the sensor field periodically. This models a common type of sensor network applications that continuously monitors the field in the long term. The objective is to minimize the period, i.e., the time the data mule takes for each travel, since it largely affects the data delivery latency. We introduce some notations for the periodic case. Let T t denote the travel time of one period. The data mule needs to stay at the base station for constant time T b to deposit the data to the base station and refuel etc. Thus the length T of one period is T t + T b. For the system to be stable, in each period of travel, the data mule needs to collect the data generated in one period. 5.1 Algorithms Processor Demand Analysis. First we present an algorithm based on processor demand analysis. This algorithm applies to the constant speed model with simple location jobs, i.e., each location job has only one feasible location interval. Let e(τ) denote the execution time of location job τ for one period. It is defined as follows: e(τ) λ(τ) R (T t + T b ), (1) where λ(τ) is the data generation rate of the node represented by τ and R is the bandwidth, both of which are known constants. Processor demand g(i) for location interval I for one period is defined as g(i) τ:i(τ) I e(τ), where I(τ) is the feasible location interval of location job τ. Let g (I) denote the processor demand for I for unit time, which is defined as follows: g (I) g(i) T t + T b = τ:i(τ) I λ(τ) R. (2)

14 14 R. Sugihara and R. K. Gupta The set of location jobs is feasible if and only if the speed v of data mule satisfies v I min I I 0 g(i) = 1 I min T t + T b I I 0 g (I), (3) where I 0 is the total travel interval. When T b > 0, we obtain the following constraint using T t = I 0 /v: ( ) I 1 v min I I 0 g (I) I 0. (4) T b For a feasible solution to exist, the following must be satisfied: I I 0 < min I I 0 g (I). (5) When this is satisfied, the maximum speed is the right hand side of (4). When this is not satisfied, it is not possible to collect data without loss. When T b = 0, we obtain the following from (3): I I 0 min I I 0 g (I). (6) Note that (6) contains neither v nor T t. What it implies is, when this is satisfied, the speed of data mule can be arbitrary. This validates the experimental observation in [Kansal et al. 2004] that the speed of data mule does not matter if the data mule travels the sensor field periodically Online Algorithm. For the variable speed model with simple location jobs and v min = 0, we can use an online algorithm almost the same as EDD-with- Stop. A data mule moves at v max while executing a job with the earliest deadline location. When a node still has data at its deadline, the data mule stops there and collects the data until the node becomes empty. Here, empty refers to the state when the buffer of a node becomes empty as the data mule collects the data from it. Note that each node keeps generating data, so it will be the case that, after the buffer of a node becomes empty, it starts to fill again while the data mule is still in the communication range. In such a case, however, the data is collected in the next period Linear Program Formulation. For the constant and variable speed models that above two approaches cannot be used (e.g., general location jobs case), we can use a linear program formulation. The formulation is almost the same as the ones for the non-periodic cases with general location jobs (Section 4.2.2, 4.3.2), but the job completion constraint is replaced with the following one: (Job completion) τ J, ( 2m i=0 p 2m ) i(τ) = i=0 z i + T b λ(τ)/r, where R is the bandwidth of communication from each node to the data mule. The right hand side is the amount of time to transmit the data generated in one period.

15 Speed Control and Scheduling of Data Mules in Sensor Networks 15 This linear program may be either infeasible or unbounded 4. When it is infeasible, it is impossible to collect all data. When it is unbounded, the speed is arbitrary Iterative Method. When there is a constraint on acceleration, we can design a heuristic algorithm based on the one for non-periodic case (Section 4.4). Specifically, we can estimate T t iteratively in the following manner: Solve the linear program for the variable speed model by ignoring the acceleration constraint. Set the result to the initial value of ˆT t, the estimate of T t. Repeat Run the heuristic algorithm with setting e(τ) = λ(τ)( ˆT t + T b )/R. Denote the travel time as T t. If T t ˆT t < ǫ, break the loop. Otherwise, update ˆT t by T t and repeat. 5.2 Feasibility Analysis For the constant speed model with simple location jobs we have discussed the feasibility conditions in Section For the variable speed model, we have a feasibility condition when v min = 0 for both simple and general location jobs cases. Let s define normalized total data generation rate α as follows: α = τ λ(τ)/r. Then the following theorem holds. Theorem 5.1. When v min = 0, there exists a finite period T if 0 α < 1. The proof is in the appendix D DMS WITH MULTIPLE DATA MULES In this section we discuss one dimensional DMS (1-D DMS) problem for multiple data mules case. 6.1 Problem Definition First we define k-dm 1-D DMS by extending the 1-D DMS problem: k-dm 1-D DMS INSTANCE: Set J of location jobs, for each location job τ J, an execution time e(τ), and a set I (k) (τ) of feasible location intervals, for each feasible location interval I (k) I (k) (τ), a release location r(i (k) ) and deadline location d(i (k) ), a start X s (k), a destination X (k) d, number of data mules K, and a constant T. QUESTION: Is there a feasible speed control plan and a feasible job schedule for K data mules satisfying T k T for all k, where T k is the travel time of k-th data mule from X (k) s to X (k) d? Depending on the mobility assumption, there may be additional constraints (e.g., maximum speed, maximum acceleration) that determine the feasibility of a speed control plan. In the rest of the paper, we focus on an optimization version of the problem, in which we minimize max k T k. 4 It may be unbounded only in the constant speed model.

16 16 R. Sugihara and R. K. Gupta For the optimization problem, we can use other objective functions as well. One example is the average travel time over all data mules k T k/n. However, it does not appropriately reflect the overall data delivery delay, for example when some of the data mules do not travel at all and have zero travel time. To avoid this, average travel time weighted by the amount of collected data ( k w kt k ) would be a more appropriate metric. However, this allows a small amount of data having a very large delay, which is also not preferable in some applications. It also has a practical issue that it is hard to be efficiently solved due to the non-linear cost function. Maximum travel time max k T k gives a guarantee on the data delivery latency, and appropriate for applications in which the maximum delivery latency is important. Although this is also a nonlinear function, it can be converted to a linear function with additional constraints, as we see later. 6.2 Constant Speed and Variable Speed First we discuss the constant speed and variable speed cases. We consider two separate cases depending on whether the paths of data mules are identical or not For Identical Paths. When paths are identical for all data mules, feasible location intervals of each job are identical as well. We define the term symmetric schedule as follows. We call a schedule is symmetric when the speed control plan and job schedule are identical in all data mules. Then we have the following theorem: Theorem 6.1. For the optimization version of k-dm 1-D DMS with identical paths for constant speed or variable speed models, there exists an optimal symmetric schedule. Proof. We show there always exists an equal or faster symmetric schedule than asymmetric one. Consider splitting the total travel interval into short location intervals by dividing at a location which is either a release or deadline location of a job. Then, without loss of generality, we can assume each data mule moves at a constant speed in each of these short location intervals. For each of these intervals, let v 1,v 2,...,v n denote the data mules speed in the location interval. Without loss of generality, we assume v 1 v 2... v n. Total amount of collected data c satisfies c i l/v i, where l is the length of the location interval. Maximum time is l/v 1. Now, consider a symmetric schedule in which each data mule moves at v = n/( i 1/v i), which is the harmonic mean of the original speed. Since n(l/v ) = i l/v i c, it is possible to collect the same amount of data as in the original schedule. Maximum time is l/v l/v 1, where the equality is satisfied if and only if v 1 =... = v n. To find an optimal schedule, we first divide the execution time of each location job equally to each data mule and then apply the optimal algorithms for single data mule case For Arbitrary Paths. When the path of each data mule is different, we formulate the problem as a linear program in the following way. The formulation is based on the same idea as the one in Section 4.3.2, but we introduce additional variable z and constraints to convert the min-max objective into a linear objective.

17 Speed Control and Scheduling of Data Mules in Sensor Networks #$ &("'% ) *+ - $./ -.,! ", Fig. 3. Example of non-optimal symmetric schedule (in acceleration-constrained case): Two location jobs have zero-length feasible location intervals with execution time e. For k-th data mule, we split the location interval [X s (k),x (k) d ] into (2m(k) + 1) location intervals [l (k) 0 (= X(k) s ),l (k) 1 ], [l(k) 1,l(k) 2 ],...,[l(k) 2m,l(k) 2m+1 (= X(k) d )] (l(k) i l (k) i+1 ), where m(k) is the number of feasible location intervals of all location jobs that are executable at this data mule, and each l (k) i is either a release location or a deadline location. Let z (k) i denote the time that the k-th data mule spends in location interval [l (k) i,l (k) i+1 ], and p(k) i (τ) denote the time it allocates to job τ in this interval. Using a variable Z to represent min-max objective, we have the following linear program: Minimize Z Subject to (Min-max objective) k, 2m (k) i=0 z (k) i Z. Note that minimizing Z is equivalent to minimizing the maximum of the left hand side over all k. (Max/min speed) k, 0 i 2m (k), v min z (k) i l (k) i+1 l(k) i variable speed model. For constant speed model, (l (k) l (k) j )/z (j) i for all i,j satisfying l (k) i+1 l(k) i > 0, l (k) j+1 l(k) j > 0. i+1 l(k) i )/z (k) i (Positive allocation time) k, 0 i 2m (k), τ J, p (k) i (τ) 0. v max z (k) i for = (l (k) j+1 (Feasible intervals) k, 0 i 2m (k), τ J, if [l (k) i,l (k) i+1 ] I(k) (τ), p (k) i (τ) = 0. (Job completion) τ J, k 2m (k) i=0 p (k) i (τ) = e(τ). (Processor demand) k, 0 i 2m (k), τ J p(k) i (τ) z (k) i. 6.3 Variable Speed with Acceleration Constraint When there is a constraint on acceleration, the problem for multiple data mules case is hard, as implied by the hardness of single data mule case.

18 18 R. Sugihara and R. K. Gupta Sort the jobs in the ascending order of the number of executable data mules. For the jobs with the same number of executable data mules, sort them by execution time in the descending order (i.e., long job first). For all jobs, from head of the list, Determine the strategy such that the maximum of current travel time is minimized. The strategy is one of the followings: Assign: Assign the job to one data mule that can execute it. Spread: Divide the job equally to all data mules that can execute it. Remove the job from the list and update the travel time of each data mule. Fig. 4. Heuristic algorithm for k-dm 1-D DMS problem with acceleration constraint One interesting observation is that, for identical paths case, symmetric schedule is not always optimal. Figure 3 shows such an example. In this example, the data mules travel from location 0 to 3l. Each of two location jobs (representing communication with two sensor nodes) have a zero-length feasible location interval at location l and 2l, respectively. All curves in location-speed graphs represent acceleration/deceleration at the maximum rate a. Since t 1 = 2 l/a and t 2 = 2 2 l/a, we have 3t 1 + e > t 1 + t 2 + e and thus the symmetric schedule is not optimal. We design a heuristic algorithm based on the idea similar to List Scheduling [Graham 1969]. Figure 4 shows the algorithm. First the jobs are sorted in decreasing order of the number of executable data mules. This is to assign the jobs that are executable at only one data mule first. In this way, there will be more freedom later in balancing the travel time of each data mule by appropriately allocating the jobs, which are executable at many data mules. The main idea of the algorithm is to assign jobs one by one to a data mule so that the maximum travel time is minimized. In addition to assigning a job to one of the data mules, we can also choose to spread the job to all data mules that can execute it by equally dividing the job s execution time, if the resulting maximum travel time is shorter than assigning the job to one data mule. When each location job is not allowed to be executed by multiple data mules, we modify the heuristic algorithm by eliminating the spread option when determining the strategy. This modified version of the algorithm is also applicable to the case without acceleration constraint. Although it is not optimal anymore when applied to constant speed or variable speed cases, it is appropriate when it is infeasible for a single node to communicate with multiple data mules. 7. PERFORMANCE EVALUATION In this section, we evaluate the performance of the proposed algorithms by numerical experiments. We first consider the variable speed with acceleration constraint model and compare the performance of the proposed heuristic algorithm with multihop forwarding and a naive speed control algorithm. Then we compare one of the optimal algorithms with a previously proposed algorithm to examine how it can improve the performance and to argue how the proposed algorithms are useful in

19 OPQ ORQ 6 8 ED C> 6 NCM H 879:;<=> 6 Speed Control and Scheduling of Data Mules in Sensor Networks 19 JL HI JKL 6 ED Fig. 5. Randomly generating test cases: (a) For comparison with multihop forwarding. Fixed nodes are aligned on line SS ; (b) For multiple data mules case. An example with two data mules (k = 2) is shown. practice. 7.1 Performance of Heuristic Algorithm in Acceleration Constrained Case For the first experiment, we compare the data delivery latency and energy consumption between data mule approach and multihop forwarding approach. Then we compare the performance of the heuristic algorithm with a naive method in case of single and multiple data mules case Method. We use Matlab for simulation experiments. Besides the heuristic algorithms, we have implemented a naive method for speed control. In the naive method, a data mule stops to collect data from each node one by one. Specifically, a data mule moves to the point on the path that is the closest to the node, collects data from the node while stopping, and moves to the next point. This is similar to the one used in [Ma and Yang 2007], though they do not assume an acceleration constraint. When there are multiple data mules, each node is assigned to the data mule whose path is the closest to it. We randomly generate test cases in the following ways for each of the two experiments. For the first experiment, we need to guarantee the connectivity between the base station and each node, since otherwise multihop forwarding approach is not feasible. For this purpose, we generate node placements as shown in Figure 5(a), in which we arrange n fixed fixed nodes on the horizontal dotted line SS, each of them are r apart, where r is the communication range. Base stations are located at S and S, assuming the data mule starts from and comes back to the base station. Then we put n n fixed random nodes at random locations in the rectangle within the distance ( 3/2)r from the horizontal line. In this way we can guarantee that each of these random nodes has at least one fixed node or the base station within its communication range and thus is reachable to the base station

20 20 R. Sugihara and R. K. Gupta Table I. Comparison with multihop forwarding: Amount of transmission is relative to the minimum possible amount. Maximum latency with asterisk is the lower bound given in [Gandham et al. 2006]. Data mule Multihop #node (n) Heuristic Naive forwarding Max. Latency (sec) (270)* 20 Transmission (max/avg) 1.0/ /5.5 Max. Latency (sec) (720)* 50 Transmission (max/avg) 1.0/ /6.0 Max. Latency (sec) (1170)* 80 Transmission (max/avg) 1.0/ /6.0 via multihop forwarding. In case of data mule approach, the data mule moves from S to S on the straight line. All the nodes are within r from the trajectory and thus the data mule can collect data from them. We set n fixed = 20, n = {20,50,80}, and r = 100[m] in the experiment. For the second experiment on multiple data mules case, we have random nodes only and no fixed nodes (Figure 5(b)). The vertical coordinate of the deployment area is [ ( 3/2)r,( 3/2)r], which is same as the first case. The path for i-th data mule is a horizontal line on the vertical coordinate of ((2i 1) 3/2k 3/2)r, where k is the total number of data mules. This setting makes each data mule cover the strip of same size. We used the same parameters as the first experiment: n = {20,50,80}, r = 100[m], and L = 2100[m]. The number of data mules is k = 1,2,3. For both experiments, we set the execution time e = 10[sec], i.e., a data mule needs 10 seconds to collect all the data from each node. Each experiment is repeated for 10 times on each case and the average is used as the result. We use the variable speed with acceleration constraint as the mobility model of data mules and set a max = 1[m/s 2 ] and v max = 10[m/s] to roughly simulate the mobility capability of a radio controlled helicopter used in [Todd et al. 2007] Comparison with Multihop Forwarding. First we compare data mule approach and multihop forwarding approach in terms of latency and energy consumption. We use one data mule in both of the heuristic and the naive algorithms for data mule approach. For a multihop forwarding algorithm, we use a simple one that each node forwards the data that it generated and received from other nodes to its neighbor node that is closest to the base station. Table I shows the maximum latency and amount of transmission in these approaches averaged for 10 different node placements for each n. Latency is measured by the total travel time in case of data mule approach. For multihop forwarding approach, we use a lower bound derived in [Gandham et al. 2006], where the authors studied the problem of minimum latency convergecast in several different network topologies. For a tree network, they gave a lower bound of max(3n k 3,N) where n k is the size of the maximum branch and N is the total number of nodes. Since we can see the network used in our experiment as a tree network whose maximum